Compressed sensing (CS) is a sampling theory that allows reconstructionof sparse (or compressible) signals from an incompletenumber of measurements, using of a sensing mechanism implementedby an appropriate projection matrix. The CS theory isbased on random Gaussian projection matrices, which satisfy recoveryguarantees with high probability; however, sparse ternary[0;-1;+1] projections are more suitable for hardware implementation.In this paper, we present a deep learning approach to obtainvery sparse ternary projections for compressed sensing. Our deeplearning architecture jointly learns a pair of a projection matrix and areconstruction operator in an end-to-end fashion. The experimentalresults on real images demonstrate the effectiveness of the proposedapproach compared to state-of-the-art methods, with significantadvantage in terms of complexity.